Session: 06-01: Machine Learning and Statistical Methods in NDE I

Paper Number: 134686

134686 - Deep Learning Method Based on Denoising Autoencoders for Temperature Selection of Guided Waves Signals 

Abstract: 

Structural Health Monitoring (SHM), and in particular Guided Wave SHM (GW-SHM), is of particular interest for high-stake industrial sectors such as aerospace, rail, or oil & gas or for which NDT inspections are complex due to a poor accessibility or large areas to inspect. Because the physics of guided waves is very complex, some inspection methods are based on a reference state, also known as baseline, which simplifies the analysis and processing of signals a posteriori.

However, comparison with a reference state is very sensitive to environmental and operational conditions (EOC). If the EOC changes, the reference state becomes obsolete and false alarms occur. To avoid this, algorithms such as Baseline Signal Stretch (BSS) are designed to compensate for a small deviation between the current state and the reference state, but are not robust to larger deviations. Coupling with Optimal Baseline Selection (OBS) can be used to limit the deviation, but this requires a large set of baseline data for each instrumented sample, often limiting the scope for industrial applications as such data is difficult to obtain due to the coupling between various EOCs.

This paper presents a method, based on denoising autoencoders, for building a model that relates guided wave signals at the current temperature to those at a selected temperature. By training the model with a sufficiently complete database of pristine and damaged signals at different temperatures for the same instrumentation on a few samples, a generic model can be obtained that can be used for all similar instrumentation. In this case, by compensating for temperature effects in the current signals to select the reference temperature, or if the current temperature is well known by selecting it for the reference signals, only one baseline is required for each sample, allowing most SHM methods to be applied to complex structures.

The method is validated on a set of signals measured on an aluminium plate with representative defects at temperatures ranging from -20°C to 50°C. The compensated signals can then be used as input to conventional SHM imaging methods, improving their results while reducing the probability of false alarms.

Presenting Author: Clément Fisher Université Paris-Saclay, CEA LIST

Presenting Author Biography: Clément Fisher was recruited at the CEA in 2019 to develop signal processing and analysis methods in the context of integrated health monitoring (Structural Health Monitoring) using guided ultrasound waves. In particular, these methods are based on machine learning and deep learning methods to automate signal analysis to perform not only detection, but also localization and characterization of defects

Authors:

Clément Fisher Université Paris-Saclay, CEA LIST
François Billon DGA, Ensta Bretagne & Université Paris-Saclay, CEA LIST
Axel Thomas Université Paris-Saclay, CEA LIST
Tom Druet Université Paris-Saclay, CEA LIST